[A-CG39] Global Carbon Cycle Observation and Analysis

[ACG39-01] Assessing central Africa tropical forest carbon using ground-based inventory and Sentinel-2 tree height deep learning

★Invited Papers

*David Roy¹, Myung Cho¹, Herve Kashongwe¹, Hankui Zhang ², Jean-Robert Bwangoy ³ (1.Center for Global Change and Earth Observations, Michigan State University, East Lansing, MI 48823, USA, 2.Geospatial Sciences Center of Excellence, Department of Geography and Geospatial Sciences, South Dakota State University, Brookings, SD 57007, USA, 3.Department of Natural Resource Management, Faculty of Agronomy, University of Kinshasa, Kinshasa, Democratic Republic of Congo)

Keywords:Tropical forest , tree height , Sentinel-2, deep learning, above ground biomass allometry

Nearly half of the world’s tropical forests are secondary forests, i.e., they are regrowing after being disturbed, and if left to regrow, provide a low-cost mechanism for carbon sequestration and an effective pathway to help mitigate climate change. Central Africa secondary tropical forests are understudied and conventional ground-based assessment of above ground biomass and carbon is extraordinarily challenging at scale. Tree height can be mapped using satellite data and related to biomass given appropriate allometric scaling factors. New advances in deep learning provide opportunities to map tree height. In this study we present a tree height map derived for 110 × 110 km of secondary forests in Mai Ndombe Province encompassing the largest Democratic Republic of the Congo (DRC) Reducing Emissions from Deforestation and Forest Degradation (REDD+) project. The recent Swin-Unet deep learning architecture, that uses an advanced attention mechanism, is trained using composited 10 m Sentinel-2 near-infrared, red and green nadir BRDF adjusted reflectance (predictor variables) and canopy height data derived from Airborne Laser Scanning (ALS) data (response variable). Spatially coherent 10 m spatial resolution tree height maps are presented and validated by comparison with independent ALS data. The results are converted into spatially explicit maps of above ground biomass and carbon estimated by considering a generalized tropical forest allometry and allometry based on tree species abundance inventoried at the REDD+ site. The implications and potential of the research for quantification of carbon sequestration in central Africa are discussed.